A Multi-Scale Feature Fusion Network for Detecting and Classifying

Apple Leaf Diseases

Assad S. Doutoum

1 a

, Recep Eryigit

2 b

and Bulent Tugrul

3 c

Computer Science Department, University of Ndjamena, Farcha, Ndjamena, Chad

Ankara University, Department of Computer Engineering, Ankara, Turkey

Ankara University, Department of Computer Science, Ankara, Turkey

Keywords:

Deep Learning, Multi-Scale CNN, Apple Leaf Diseases.

Abstract:

Early detection and identiﬁcation of leaf diseases reduce expenses and increase proﬁts. Thus, it is essential

for producers to be aware of the symptoms and indications of these leaf diseases and take the necessary pre-

ventative measures. Early diagnosis and treatment can also help prevent the disease from spreading to healthy

plants. For successful disease control, regular inspections of orchards are essential. As well as being costly and

time-consuming, traditional methods require a great deal of labor. However, the use of modern technologies

and methods such as computer vision will both increase successes and reduce costs. Deep learning methods

can be used to detect and classify diseases, as well as predict the likelihood of them occurring. Though, a

particular CNN architecture may focus on a subset of features, while another may discover other additional

features not extracted from the dataset. Robust classiﬁcation models should be developed that perform con-

sistently well when different environmental factors such as light, angle, background and noise vary. To solve

these challenges, this study proposes a multi-scale feature fusion network (MFFN) that combines features

from different scales or levels of detail in an image to improve the performance and robustness of classiﬁca-

tion models. The proposed method is evaluated on a publicly available dataset and is shown to improve the

performance of the original models. Four branches applied to CNN architectures were simultaneously trained

and merged to accurately classify and predict infected apple leaves. The merged model was able to detect

infected leaves with a high degree of accuracy, signiﬁcantly through the combined models. The merged model

was able to accurately predict the unhealthy apple leaves with a 99.36% on the training accuracy, 98.90% on

the validation accuracy, and 98.28% on the test accuracy. The results show that combining the models is an

effective way to increase the accuracy of predictions under volatile conditions.

1 INTRODUCTION

Plant leaf diseases severely affect the quality and pro-

ductivity of agricultural products. The agricultural in-

dustry loses millions of dollars every year due to yield

loss and unnecessary or misuse of pesticides (Alsayed

et al., 2021). It is therefore crucial to detect and diag-

nose plant diseases early to prevent their spread and

minimize their impacts. Several pathogens such as

Black Rot disease, Cedar Rust disease, and Scab dis-

ease may affect apple leaves (Su et al., 2022; Nand-

hini and Ashokkumar, 2022). Traditional methods

of detecting plant diseases require specialized knowl-

https://orcid.org/0000-0003-4822-1876

https://orcid.org/0000-0002-4282-6340

https://orcid.org/0000-0003-4719-4298

edge and expertise in plant pathology. These meth-

ods are often time-consuming and require the use of

expensive equipment, making them inaccessible to

those without adequate budgets. These limitations

hinder the ability to detect and monitor diseases on

small farms and in low-income countries where re-

sources and experts are scarce. Consequently, ad-

vanced technologies and automated systems are re-

quired to detect plant sicknesses at an early stage.

With the help of such technologies it is possible

to detect, classify and diagnose plant diseases more

quickly and accurately. Thanks to these methods,

agricultural enterprises can save time and money and

gain a competitive advantages over their competitors

on both national and international markets.

Various machine learning and deep learning meth-

ods have been used to detect and classify plant ill-

Doutoum, A. S., Eryigit, R., Tugrul and B.

A Multi-Scale Feature Fusion Network for Detecting and Classifying Apple Leaf Diseases.

DOI: 10.5220/0013447700003967

In Proceedings of the 14th International Conference on Data Science, Technology and Applications (DATA 2025), pages 13-20

ISBN: 978-989-758-758-0; ISSN: 2184-285X

nesses (Barbedo, 2018; Doutoum and Tugrul, 2023).

It is suggested that deep learning methods shorten

the time-consuming and cost-efﬁcient process of fea-

ture extraction in image processing (Yan et al., 2020).

The models obtained through deep learning meth-

ods demonstrate signiﬁcantly improved accuracy and

robustness compared to traditional machine learning

methods. In traditional machine learning, prediction

or classiﬁcation models are trained using manually

extracted features. In contrast, deep learning mod-

els have the ability to automatically extract features

from images and build the model on them. As a re-

sult, manual feature extraction is no longer necessary,

increasing the efﬁciency and reducing the cost of clas-

siﬁcation methods. Furthermore, deep-learning algo-

rithms have the potential to recognize complex corre-

lations and patterns as well as various high-level ele-

ments that are invisible to humans. As a result, they

are more suitable for tasks that require more domain-

speciﬁc knowledge and experience.

Feature fusion uses multiple scales or levels of de-

tail in computer vision and image processing to in-

tegrate facts into feature representation to enable a

better understanding of visual elements (Elizar et al.,

2022). As a result, they are more resistant to changes

in object size, orientation, and scale, because charac-

teristics at various sizes can give complementing in-

formation. This may rise item recognition and clas-

siﬁcation accuracy, as well as the capability to dis-

tinguish things in a more problematic context. There

are many ways of accomplishing multi-scale feature

fusion in computer vision. One possible approach is

to use distinct Convolutional Neural Network (CNN)

architectures to extract features of varying sizes and

aspects and then efﬁciently integrate them. Integrat-

ing detailed features from small scales with contex-

tual knowledge from large scales yields more com-

prehensive and robust models.

The contributions of this article can be summa-

rized as follows: By proposing an architecture based

on multi-scale CNN for apple leaf disease classiﬁca-

tion and prediction, we improve the overall perfor-

mance of the model by allowing it to better under-

stand the spatial structure of the images, leading to

more accurate predictions. Our proposed model is

based on parallel training of big datasets, is less com-

plex, and has accurate results compared with other

deep learning models (Yan et al., 2020; G

und

uz and

Yılmaz G

und

uz, 2022). According to the compari-

son dataset, our model achieved the highest accuracy

result of 98.28% on the test set.

2 METHODS AND MATERIALS

The following section provides detailed information

about the Apple data source. Additionally, the pre-

processing procedures applied to the dataset and the

training process of the proposed multi-scale feature

fusion network (MFFN) will be explained.

2.1 Dataset

The dataset utilized in this study consists of images

with healthy and unhealthy apple leaves. As can be

seen from Table 1, there are four main classes within

the dataset. The ﬁrst class contains leaf images of

Black Rot disease. The second class contains images

of Cedar Rust disease. The third and fourth classes

contain images of healthy leaves and Scab disease re-

spectively. In total, there are 10,173 images of apple

leaves in the dataset (Hughes and Salath

e, 2015). The

class distribution of the apple dataset is shown in Ta-

ble 1.

2.2 Data Pre-Processing

Image pre-processing and validation is an essential

step to calibrate the images before being trained by

the deep learning model (Divakar et al., 2021). Deep

learning classiﬁers require huge datasets to overcome

over-ﬁtting problems and obtain better precision level

of achievement. It is therefore necessary to perform

an augmentation techniques in order to expand the

dataset from the one that currently exists. This re-

quires changes such as resizing, zooming, rotating, or

adjusting the color of the images. Nevertheless, it is

necessary to split the dataset into training, validation,

and testing sets (G

und

uz and Yılmaz G

und

uz, 2022).

Using the image data augmentation techniques, we

horizontally shift the dataset’s images.

2.3 Dataset Augmentation Techniques

Common data augmentation techniques include ﬂip-

ping, rotating, and scaling images to create new vari-

ations. Other methods involve adjusting brightness,

contrast, and adding noise to enhance the dataset fur-

ther. By expanding the dataset, it allows models to

learn features more robustly and generalize better to

unseen data. Additionally, it introduces variability

that can make the model more resilient to noise and

variations in real-world data. This leads to improved

accuracy and performance across diverse scenarios.

Common data augmentation techniques include ﬂip-

ping, rotating, and scaling images to create new vari-

ations. Other methods involve adjusting brightness,

DATA 2025 - 14th International Conference on Data Science, Technology and Applications

Table 1: Distribution of the apple dataset.

Disease type Black Rot Cedar Rust Healthy leaf Scab Total

Number of images 2352 2792 2510 2519 10173

Figure 1: Sample of images after the augmentation pro-

cesses.

contrast, and adding noise to enhance the dataset fur-

ther (Baranwal et al., 2019; Su et al., 2022). Table 2

displays the augmentation details. Figure 1 shows a

sample of images after the augmentation processes.

Table 2: Augmentation parameters.

Parameter Value

Rotation 40

Width 0.2

Height 0.2

Horizontal ﬂipping True

Vertical ﬂipping True

Shear 0.2

Zoom 0.2

2.4 Data Splitting

The dataset used in this study contains a total of

10173 images. The dataset is divided into three fold-

ers: the training and the validation sets contain 6353

and 2014 images respectively. The remaining 1806

images are assigned to the test set. The validation

set is allocated to assess the model’s generalization

ability. If over-ﬁtting is observed, hyper-parameters

such as learning rate, epoch size, or number of lay-

ers can be ﬁne-tuned to enhance the model’s perfor-

mance. This process of ﬁne-tuning hyper-parameters

based on validation results help to identify the opti-

mal conﬁguration that maximizes the model’s perfor-

mance on unseen test data. The apple dataset is di-

vided into four classes, and each class contains one

apple disease (Yadav et al., 2022).

2.5 Multi-Scale CNN Model Structure

The improved model architecture is based on the

multi-scale CNN model, which is shown in Figure 2.

Our network consists of one input image shape with a

size of 224×224×3 divided into four branches. Each

CNN branch has one convolution layer (64×64) fol-

lowed by Batch Normalization layers (BN) and Max-

Pooling layers. After each Batch Normalization we

use ReLu as an activation function. We add separa-

ble layers before concatenating the four branches net-

works. Following the merging of the CNN networks,

two dense layers are added to the network. Table 3

displays the description of the Multi-scale CNN net-

work architecture.

Table 3: Description of MFNN architecture.

Layer CNN (Input=224 × 224 × 3)

Conv1 1 × 1

64 × 64

MaxPool 2 × 2

Conv2 1 × 1

64 × 64

MaxPool 2 × 2

Conv3 1 × 1

64 × 64

MaxPool 2 × 2

Conv4 1 × 1

64 × 64

MaxPool 2 × 2

Concatenate conv1, conv2 , conv3 , conv4

Conv5 3 × 3

64 × 64

MaxPool 2 × 2

Conv6 3 × 3

64 × 64

MaxPool 2 × 2

Conv7 3 × 3

128 × 128

MaxPool 2 × 2

Dense -

2.6 Experimental Environment and

Parameters

For the purpose of training, our proposed model for

classiﬁcation and detection of Apple leaf diseases,

used TensorFlow/Keras framework, and Python 3.11

A Multi-Scale Feature Fusion Network for Detecting and Classifying Apple Leaf Diseases

Figure 2: MFFN architecture.

version. Table 4 shows the hardware and software

details of the experimental environment. The Apple

dataset was divided into three parts: training, valida-

tion, and testing sets. There are multiple parameters

applied to the training and validation dataset. Table

5 shown the parameters utilized during training and

validation.

3 EXPERIMENTAL RESULTS

AND ANALYSIS

This section evaluates the performance of our pro-

posed MFFN model in the classiﬁcation and detection

of apple leaf diseases. The experimental evaluates and

analyzes confusion metrics and compares the module

with other existing models. To evaluate the efﬁciency

of the proposed model we trained on the apple leaf

dataset, which consists of four distinct classes. The

percentage of true guesses in prediction results com-

pared to all guesses is known as accuracy rate. Recall

rate is the percentage that a category is accurately pre-

dicted in all actual values, whereas precision rate is

the percentage that a category is correctly forecasted

in all prediction outcomes.

3.1 Evaluation Metrics

The confusion matrix is used to evaluate the achieve-

ments of our model. A confusion matrix is presented

in this section for a thorough analysis. The confu-

sion matrix’s rows correspond to the actual category

labels, while its columns correspond to the labels that

the model predicted. Through the number of cases of

True Positive (TP), False Positive (FP), True Negative

DATA 2025 - 14th International Conference on Data Science, Technology and Applications

Table 4: Experimental hardware and software.

Name Model

Framework TensorFlow/Keras

GPU NVIDIA RTX A4000

CPU Intel(R) Xeon(R) W-2235 CPU @ 3.80GHz, 3792 MHz

OS Win 11 Pro

Ram 32 GB

Python 3.11

Table 5: Experimental parameters setting.

Parameter Name Parameter setting

Learning rate 0.001

Batch size 32

Optimizer Adam

Activation function SoftMax

Beta1 0.9

Beta2 0.999

Amsgrad False

Iteration epochs 100

(TN), and False Negative (FN), the confusion matrix

produces four assessment metrics: accuracy, recall,

speciﬁcity, and F1 score. The formula for each metric

is given below.

Precision =

T P

T P + FP

(1)

Recall =

T P

T P + FN

(2)

F1 − score = 2 ×

Precision × Recall

Precision + Recall

(3)

It is evident from the confusion matrix displayed

in Figure 3 that the majority of classiﬁcation errors

occur in scenarios where there is a distinction be-

tween cedar rust and scab. One of the main reasons

for this is that the complex background context of the

data set facilitates errors in judgment. The early dis-

ease spots in the Scab and Cedar rust dataset are pri-

marily small-scale features with intricate background

settings. Another reason for that is the texture of

the affected areas might appear irregular or spotted,

which is a common feature between the two diseases.

Additionally, the impact of nearby leaves that fall into

different health categories results in classiﬁcation er-

rors and insufﬁcient reliable information about the ﬁ-

nal marks that the model obtains. Nonetheless, the

model’s overall effect still complies with the require-

ments of the leaf disease detection scenario. Figure

4 displays precision, recall, and F1 scores for each

class.

Figure 3: Confusion matrix of the MFFN.

Figure 4: Classiﬁcation performance of the MFFN.

Table 6: Training and validation results.

Metrics Training Validation

Accuracy 99.36% 98.90%

Loss 0.1072 0.1140

3.2 Experimental Results

The MFFN model was trained with an apple dataset

for four classes. These classes contain different types

A Multi-Scale Feature Fusion Network for Detecting and Classifying Apple Leaf Diseases

(a) Training and validation accuracy (b) Training and validation loss

Figure 5: Training and validation results.

of apple leaf diseases including Black Rot disease,

Cedar Rust disease, and Scab disease. Our model was

trained for 100 epochs to achieve the best accuracy in

training and validation. The outcomes of the assess-

ment highlighted areas in need of development, of-

fered insightful information about the model’s func-

tionality, and improved it even further. Finally, the

performance review showed that our suggested model

is a strong solution in its ﬁeld and very successful at

accomplishing the aims it set out to do.

The preciseness of the training and validation of

the suggested model demonstrated remarkable accu-

racy. The model performed well, as evidenced by the

average training and validation accuracy of 99.36%

and 98.90%, respectively. Furthermore, it was noted

that the model’s thoroughness was further supported

by the training and validation losses, which were

0.1072 and 0.1140, respectively. Figure 5 and Table

6 indicate that our proposed model achieved the best

performance during training on the apple plant village

dataset for apple diseases classiﬁcation.

3.3 Model Performance

The efﬁciency of the proposed MFFN model has been

compared with other multi-scale CNN models to con-

ﬁrm its performance. Each model is trained while

maintaining the aforementioned experimental proce-

dure and experimental setting, and the outcomes are

shown in Table 7. Our model demonstrates an excel-

lent stability, as demonstrated by its high initial iden-

tiﬁcation accuracy, low error rate, and rapid conver-

gence throughout both training and validation.

Improvements have been observed in the disease

classiﬁcation of Apple leaves using deep learning

techniques. The suggested approach represents a ma-

jor advancement in the detection and classiﬁcation of

Apple leaf diseases, enabling precise and effective re-

sults. It is worthwhile to note that the recommended

model can be made even better by adding more data

sources, adjusting hyperparameters, and investigating

cutting-edge deep learning architectures.

3.4 Comparison Between Different

Approaches and Our Proposed

Model

To examine the effectiveness and performance of the

proposed MFFN, we compare our model with 10

other techniques that are used to classify Apple leaf

diseases. The outcomes are contrasted with those of

other pre-trained models, such as the built exclusively

model. Table 8 provides an overview of the ﬁndings.

From the table we ﬁnd out that our suggested method

performs signiﬁcantly better than pre-trained mod-

els and built models. From the comparison between

the techniques and MFFN in Table 8, we observed

that our provided MFFN model has an accuracy rate

higher than the other architectures in the ﬁeld of Ap-

ple leaf diseases classiﬁcation.

4 CONCLUSION AND FUTURE

WORKS

Apple leaf diseases can signiﬁcantly affect produc-

tivity and quality by causing early fruit drops and

compromising the appearance, size, and ﬂavor of the

fruit. This can lead to ﬁnancial losses for farmers

and a reduction in the overall quality of apple con-

sumption. As a result, it can lead to the economic

decline of the agricultural industry. Therefore, early

diagnosis of apple leaf disease and preventative mea-

DATA 2025 - 14th International Conference on Data Science, Technology and Applications

Table 7: MFFN performance compared with various multi-scale CNN.

Reference Plant Methodology Accuracy

Ren et al. (Ren et al., 2023) Apple Multi-scale parallel fusion 97.20%

Ahmed et al. (Ahmed et al., 2022) Apple Multi-Contextual Feature Fusion Network 90.86%

Luo et al. (Luo et al., 2021) Apple Multi-scale extraction of disease features 94.23%

Suo et al. (Suo et al., 2022) Grape Multi-scale fusion module 95.95%

Tian et al. (Tian et al., 2022) Apple Multi-scale feature extraction 83.19%

Proposed MFFN Apple Our proposed model 98.28%

Table 8: Comparison between different approaches and our proposed model on Apple leaf diseases.

Reference Methodology Plant Accuracy

Zhang et al. (Zhang et al., 2021) ResNet34 Apple 93.76%

Sangeetha et al. (Sangeetha et al., 2022) VGG16 Apple 93.30%

Zhao et al. (Zhao and Huang, 2023) CBAM-ENetV2 Network Apple 97.49%

Alsayed et al. (Alsayed et al., 2021) ResNetV2 Apple 94.00%

Sheng et al. (Sheng et al., 2022) MobileViT-based Apple 96.76%

Rehman et al. (Rehman et al., 2021) Mask RCNN Apple 96.60%

Chakraborty et al. (Chakraborty et al., 2021) SVM Apple 96.00%

Assad et al. (Assad et al., 2023) AppleNet Apple 96.00%

Chen et al. (Chen et al., 2023) ResNet Apple 97.78%

Yu et al. (Yu et al., 2022) ResNet-50 Apple 95.70%

Proposed MFFN MFFN Apple 98.28%

sures can reduce the economic loss. Multiscale CNNs

can enhance apple leaf disease classiﬁcation by ana-

lyzing local and global features. They accommodate

size variability in lesions, handle texture changes, and

make the model more robust to real-world variations.

Therefore, we utilized deep learning based on a multi-

scale CNN approach to identify apple leaf diseases

with a high test accuracy rate of 98.28%. The demon-

strated method has shown signiﬁcant and robust re-

sults in evaluation metrics such as precision, recall,

and F1-score. After applying a deep learning model

to the apple leaf dataset to detect apple leaf diseases

in their earlier stages and achieving promising re-

sults, this simulation can be deployed as a web- or

mobile-based application. Such applications will help

farmers detect apple leaf diseases in real life. In fu-

ture works, our proposed multi-scale CNN technique

can be implemented to classify different plant species.

Besides, it could be employed to identify other apple

leaf diseases, making it easier for farmers and produc-

ers to categorize multiple diseases. Overall, the im-

plementation of the proposed model may help farmers

make better-informed decisions.

REFERENCES

Ahmed, M. R., Ashraﬁ, A. F., Ahmed, R. U., and Ahmed,

T. (2022). Mcffa-net: Multi-contextual feature fusion

and attention guided network for apple foliar disease

classiﬁcation. In 2022 25th International Conference

on Computer and Information Technology (ICCIT),

pages 757–762. IEEE.

Alsayed, A., Alsabei, A., and Arif, M. (2021). Classiﬁca-

tion of apple tree leaves diseases using deep learning

methods. International Journal of Computer Science

& Network Security, 21(7):324–330.

Assad, A., Bhat, M. R., Bhat, Z., Ahanger, A. N., Kun-

droo, M., Dar, R. A., Ahanger, A. B., Dar, B., et al.

(2023). Apple diseases: detection and classiﬁcation

using transfer learning. Quality Assurance and Safety

of Crops & Foods, 15(SP1):27–37.

Baranwal, S., Khandelwal, S., and Arora, A. (2019). Deep

learning convolutional neural network for apple leaves

disease detection. In Proceedings of International

Conference on Sustainable Computing in Science,

Technology & Management (SUSCOM-2019), pages

260–267.

Barbedo, J. G. (2018). Factors inﬂuencing the use of deep

learning for plant disease recognition. Biosystems En-

gineering, 172:84–91.

Chakraborty, S., Paul, S., and Rahat-uz Zaman, M. (2021).

Prediction of apple leaf diseases using multiclass sup-

port vector machine. In 2nd International Conference

on Robotics, Electrical and Signal Processing Tech-

niques (ICREST), pages 147–151.

Chen, Y., Pan, J., and Wu, Q. (2023). Apple leaf disease

identiﬁcation via improved cyclegan and convolu-

tional neural network. Soft Computing, 27(14):9773–

9786.

Divakar, S., Bhattacharjee, A., and Priyadarshini, R. (2021).

Smote-dl: A deep learning based plant disease detec-

tion method. In 6th International Conference for Con-

vergence in Technology (I2CT), pages 1–6.

A Multi-Scale Feature Fusion Network for Detecting and Classifying Apple Leaf Diseases

Doutoum, A. S. and Tugrul, B. (2023). A review of leaf

diseases detection and classiﬁcation by deep learning.

IEEE Access, 11:119219–119230.

Elizar, E., Zulkiﬂey, M. A., Muharar, R., Zaman, M. H. M.,

and Mustaza, S. M. (2022). A review on multiscale-

deep-learning applications. Sensors, 22(19):7384.

und

uz, H. and Yılmaz G

und

uz, S. (2022). Plant disease

classiﬁcation using ensemble deep learning. In 2022

30th Signal Processing and Communications Applica-

tions Conference (SIU), pages 1–4.

Hughes, D. P. and Salath

e, M. (2015). An open ac-

cess repository of images on plant health to en-

able the development of mobile disease diagnostics

through machine learning and crowdsourcing. CoRR,

abs/1511.08060.

Luo, Y., Sun, J., Shen, J., Wu, X., Wang, L., and Zhu, W.

(2021). Apple leaf disease recognition and sub-class

categorization based on improved multi-scale feature

fusion network. IEEE Access, 9:95517–95527.

Nandhini, S. and Ashokkumar, K. (2022). An automatic

plant leaf disease identiﬁcation using densenet-121 ar-

chitecture with a mutation-based henry gas solubility

optimization algorithm. Neural Computing and Ap-

plications, 34(7):5513–5534.

Rehman, Z. U., Khan, M. A., Ahmed, F., Dama

sevi

cius, R.,

Naqvi, S. R., Nisar, W., and Javed, K. (2021). Recog-

nizing apple leaf diseases using a novel parallel real-

time processing framework based on mask rcnn and

transfer learning: An application for smart agriculture.

IET Image Processing, 15(10):2157–2168.

Ren, H., Chen, W., Liu, P., Tang, R., Wang, J., and Ge, C.

(2023). Multi-scale parallel fusion convolution net-

work for crop disease identiﬁcation. In 2023 8th In-

ternational Conference on Cloud Computing and Big

Data Analytics (ICCCBDA), pages 481–485. IEEE.

Sangeetha, K., Rima, P., Kumar, P., Preethees, S., et al.

(2022). Apple leaf disease detection using deep learn-

ing. In 6th International Conference on Computing

Methodologies and Communication (ICCMC), pages

1063–1067.

Sheng, X., Wang, F., Ruan, H., Fan, Y., Zheng, J., Zhang,

Y., and Lyu, C. (2022). Disease diagnostic method

based on cascade backbone network for apple leaf

disease classiﬁcation. Frontiers in Plant Science,

13:994227.

Su, J., Zhang, M., and Yu, W. (2022). An identiﬁcation

method of apple leaf disease based on transfer learn-

ing. In 7th International Conference on Cloud Com-

puting and Big Data Analytics (ICCCBDA), pages

478–482.

Suo, J., Zhan, J., Zhou, G., Chen, A., Hu, Y., Huang, W.,

Cai, W., Hu, Y., and Li, L. (2022). Casm-amfmnet: A

network based on coordinate attention shufﬂe mech-

anism and asymmetric multi-scale fusion module for

classiﬁcation of grape leaf diseases. Frontiers in Plant

Science, 13:846767.

Tian, L., Zhang, H., Liu, B., Zhang, J., Duan, N., Yuan, A.,

and Huo, Y. (2022). Vmf-ssd: A novel v-space based

multi-scale feature fusion ssd for apple leaf disease

detection. IEEE/ACM Transactions on Computational

Biology and Bioinformatics.

Yadav, A., Thakur, U., Saxena, R., Pal, V., Bhateja, V.,

and Lin, J. C.-W. (2022). Afd-net: Apple foliar dis-

ease multi classiﬁcation using deep learning on plant

pathology dataset. Plant and Soil, 477(1):595–611.

Yan, Q., Yang, B., Wang, W., Wang, B., Chen, P., and

Zhang, J. (2020). Apple leaf diseases recognition

based on an improved convolutional neural network.

Sensors, 20(12):3535.

Yu, H., Cheng, X., Chen, C., Heidari, A. A., Liu, J., Cai, Z.,

and Chen, H. (2022). Apple leaf disease recognition

method with improved residual network. Multimedia

Tools and Applications, 81(6):7759–7782.

Zhang, D., Yang, H., and Cao, J. (2021). Identify apple leaf

diseases using deep learning algorithm. arXiv preprint

arXiv:2107.12598, abs/2107.12598.

Zhao, G. and Huang, X. (2023). Apple leaf disease recog-

nition based on improved convolutional neural net-

work with an attention mechanism. In 5th Interna-

tional Conference on Natural Language Processing

(ICNLP), pages 98–102.

DATA 2025 - 14th International Conference on Data Science, Technology and Applications